A Reinforcement Learning Approach to Control.

Abstract : Active perception strategies are necessary for goal driven allocation of available resources to improve relevant information acquisition and optimize overall system performance. In addition to being both goal and data driven, these strategies must also account for the fact that information acquisition is inherently a partially observable Markov decision problem. This report describes an efficient, scalable reinforcement learning approach to the control of autonomous active vision that also satisfies the more stringent requirements of foveal machine vision. Foveal vision offers images with both wide field of view, useful for rapid detection, and a high acuity zone, useful for accurate recognition, without the overhead and errors inherent in dynamic registration of data from multiple sensors. However, space variant data acquisition inherent with foveal retinotopologies necessitates deployment of refined intelligent gaze control techniques. This report first lays a theoretical foundation for reinforcement learning. It then introduces the SARSA algorithm in conjunction with history augmentation as an effective learning control method for visual attention. The system is shown to perform well in both high and low SNR ATR environments. Reinforcement learning coupled with history features appears to be both a sound foundation and a practical scalable base for gaze control.